The continuous work features announced in the period, with agents executing long tasks without supervision at each step, bring a requirement that usually arrives after the first incident: a record of what was done.
The need is simple to state. Code review assumes you can reconstruct what changed and why. When part of the changes happen with nobody watching, that reconstruction depends entirely on what was recorded.
Useful logging here isn't the technical execution trace, which usually exists. It's the decision record: what was tried, what was discarded, which file was changed and with what justification.
Without it, review looks only at the final result, which works when it's right and fails exactly when it matters most, which is when something came out different from expected.
The practical recommendation accompanying this kind of feature precedes its use: define the agent's reach in the repository, require confirmation on irreversible actions, and check that the log can answer who authorised what.